The clunky early stack point is something I think about a lot. My setup has about 40 shell scripts, 3 Python watchers, and a folder structure that made sense 6 months ago but is now held together with duct tape. None of it is pretty. All of it runs.
The Make quirks you mention - I hit the same kind of friction with cron job sequencing. Took 3 weeks to get reliable execution order. But the agent is now doing work I'd otherwise skip entirely. The data and reputation moat argument is more compelling to me than the technology moat.
The tools will commoditize. What compounds is the operating knowledge of what actually works at your specific scale.